Adapting to AI penetration testing orchestrators: The Modern Security Model

Over the past month, the number of AI-powered penetration testing frameworks has surged globally, with thousands of IP addresses now running these tools. The original framework, ARTEX, was taken down on October 8, but its removal triggered a rapid proliferation of over 12,000 repositories and hundreds of forks. Daily requests for these tools have increased 30-fold since October 1, with some attacks chaining stolen credentials for lateral movement. Defenders are struggling to keep pace with the volumetric and autonomous nature of these threats.

ARTEX and its derivatives enable AI agents to automate stages of the Cyber Kill Chain, mapping environments and executing attacks at inference speed. These tools can orchestrate multiple agents in parallel, turning compromises into optimization problems that AI excels at solving. As of November 12, ARTEX and its forks account for over 60% of known AI penetration testing orchestrators, with traffic growing more than tenfold since October 8.

Incident response investigations by Cloudflare’s Cloudforce One team revealed attackers exploiting static, non-expiring API tokens in CI/CD pipelines to pivot into customer environments. In one case, an ARTEX-derived framework accessed a compromised API token with no IP restrictions or rate limiting, originating from a third-party hyperscaler. The team observed over 69 requests per minute and 400 database dump requests in an hour, alongside attempts to validate session tokens for other applications.

Cloudflare advocates adopting the Modern Security Model, which emphasizes perimeter defenses, behavioral analysis, and credential hygiene. The model compares cyber defense to securing an estate, where volumetric attacks are filtered at the perimeter and AI-driven intruders are identified by unnatural behaviors. Web Application Firewalls and Bot Defense tools are critical in detecting and mitigating these automated threats before they escalate.